Orchestrata Ultimate › Sessions › Session 269

Session 269

‹ previous · next › · timeline

Prose pending.

Purpose

test-run-context hook: a foreign CI run was counted as this checkout's own authoritative run, failing a ship

Objectives

• Identify a run by checkout ROOT and owner UID instead of the colliding project NAME
• Keep a foreign run REPORTED as contention — it saturates the same box — while not counting it as ours
• Prove it with real processes read out of real /proc, no faked process tree
• Ship: regenerate the plugin delivery path, deploy, and verify against the live CI job

Work log

• 2026-08-26 17:07 Escalation logic: an untested changed file escalated to the full suite, which runs ZERO tests for that file — maximum cost for zero verification
  • affected_tests.py now reports untested files instead of escalating; run_tests.py exits 1 and names the test to write
  • Added the stem convention (scripts/X.py binds to tests/scripts/test_X.py) so a script with a test binds instead of escalating
  • Wrote tests/scripts/test_apply_docstring_traces.py (20 tests), which also binds the generator and removes its escalation trigger permanently
  • Measured payoff: 1835 modules and roughly 8 min became 11 modules and 55s, no escalation. Shipped, DEPLOYED prod at 368690f8e5ef
• 2026-08-26 17:07 Foreign-run identity: a GitLab CI job in /builds/MSamy counted as one of THIS checkout's authoritative runs
  • _project_of returns the repo DIRECTORY NAME, identical across checkouts, so CI was indistinguishable from the local run
  • Under ship that made two authoritative runs where the collapse invariant expects one, and failed a ship on 2026-08-26
  • Added _repo_root_of and _here_root for path identity, plus _uid_of reading the Uid line of /proc/PID/status for owner identity
  • Shipped b1461bcaa; deployed to ~/.claude (139 files verified) and prod
• 2026-08-27 01:02 P1 backlog staleness sweep: verified all 21 P1 items against the live tree
  • Four were already fixed and never closed: AST refs alias blind spot (refs now tags [ALIAS] and emits an alias-scan WARN; unused --ref-source defaults project-wide), autoformat stripping used imports (_NEVER_AUTOFIX F401/F841 passed to ruff --unfixable), human override inert beyond Article I (5caacb60d; constitutional.pl:108 and intent_policy.pl:147 consume it), plans/analysis untracked (9 files tracked, RUL-030 gate green).
  • Four re-scoped against measurement: pyright 410 errors ALL in tests (342) and claude (68), src and scripts at ZERO; Phase-3 codegen enums lane BUILT but ORPHANED (specs/codegen/enums.json has 0 non-test refs, no gate drives gen-enum --check); capability integration down to the brief-pack arm alone, P1 to P2; traceability re-measured at 2224/5561 untraced with orphans WORSE at 4711, plus 639.
  • Nine confirmed live with fresh evidence, including: constitutional evidence predicates declared in the .pl rulebase but with ZERO Python references so production never asserts them; no cancel symbol in either pipeline module so F-4 cooperative cancel is unstarted; .coverage carries exactly one context and it is empty, confirming T1b has no per-test contexts.
  • Removed at Mohamed's instruction, not resolved: the DEEPSEEK/OPENROUTER rotation item, both Claude credential-mirror items, and the gateway.rotate_token admin RPC.
• 2026-08-27 03:18 Pyright debt: 410 errors across tests/ and claude/ cut to 60; src and scripts held at zero
  • Root causes found and fixed, not suppressed. (1) 22 hooks carried a try/except import of orchestrata.core.traceability that has NEVER resolved -- importlib raises ModuleNotFoundError, so the no-op fallback was always the live branch -- and each one ran sys.path.insert(0, <ancestor>) at import time, the stdlib-shadowing hazard. trace-audit proved harvesting is AST-based (all 8 decorators in stop_auditor.py are read despite the failed import), so the block was deleted with zero RTM loss.
  • (2) pyrightconfig had no extraPaths, so 43 runtime-resolvable imports (session.models, session_cli via claude/tools) were unresolved. Adding claude/tools alone made it WORSE, 388 to 450: without src listed too, orchestrata_ultimate resolved as an installed library and 90 reportPrivateImportUsage appeared. extraPaths [src, claude/tools] is the correct pair.
  • (3) 98 errors sat in tests/generated_recovery + tests/_generated_broken, which pytest addopts --ignore and therefore never executes; excluded from pyright to match collection, the same treatment check_async_await.py already applies. (4) typings/codemod_fast/init.pyi declared only the 4 public functions, so the 28 assertions exercising the pure-Python fallbacks (_py_csr_build, _HAS_NATIVE, ...) were unknown-attribute errors against a module that plainly has them; stub completed.
  • (5) The ModuleType family, 98 errors: tests load their subject by file path and then set module globals, so ModuleType is a false annotation promising an attribute surface the object does not have. Retyped to Any at the loader function and at module-level module_from_spec bindings. (6) GraphCompactor.init(self, driver=None) had no annotation, so self._driver was None for the whole class and every self._driver.session() was an error; collapsed to one annotated assignment.
  • CORRECTION: I repointed context_graph's orchestrata.<mod> imports at context_graph.<mod> on the evidence that orchestrata does not resolve from the repo root. Wrong -- test_context_graph_init_coverage.py loads the package UNDER the name orchestrata via spec_from_file_location with submodule_search_locations, so orchestrata.compactor is correct and my change broke the package. Reverted; those 7 imports now carry a documented per-line pyright ignore, the narrow case where a name exists only at runtime.
  • Verified: 129/129 affected test modules pass, ruff and black clean, plugin regenerated so the fix reaches the delivery path.

Artifacts

Decisions

• Identify a run by BOTH checkout root path and owner uid, never by project name — The two signals fail at different boundaries. /proc/PID/cwd is a symlink only its owner may resolve, so across users the root degrades to empty and a root-only rule never fires — that is the exact CI case that failed the ship, and it rendered as 'pid 1915816 dot ?'. /proc/PID/status is world-readable, so uid survives that boundary. Root still catches the same-user case: a worktree, a clone, a deployment.
• Consult each ground only when its input is KNOWN; an unreadable uid or unidentifiable root is never assumed foreign — The error is one-way. Wrongly calling a run foreign moves a real contender into the quieter bucket and hides it; wrongly calling it ours only over-warns. Fail toward the visible answer.
• Make the rule a named predicate _is_foreign(record, here, here_uid) taking the viewpoint as parameters — Spawning a process under another uid needs sudo, and a sudo-dependent test either gets skipped (forbidden here) or fails on other machines. With the viewpoint as a parameter the record stays REAL — produced by _runs() off real /proc — and only the question changes. A rule you cannot ask from another viewpoint cannot be tested from one.
• Orphan the foreign test process deliberately instead of spawning a plain child — A CI job is not a descendant of this session, and the hook folds descendants of an authoritative run into it as workers. A plain child would have been attributed to whichever runner was executing the suite and never reached the branch under test — the test would have passed under bare pytest and silently proved nothing under run_tests.py.

Achievements

• Session archived by orchestrata stop

Related

• Session log
• Session timeline
• Orchestrata Ultimate — project index
• Workflow